Curiosity-Driven MARL Method for Dynamic Waveform Games Between Intelligent Radar and Jammer

The competition between radars and jammers is becoming increasingly dynamic and intelligent in modern electronic warfare. Different levels of the agents considerably influence the behaviors of the opponents, which renders the radar–jammer game a dynamically evolving competition with unstable interactions. The feasibility of leveraging multiagent reinforcement learning approaches under the above gaming conditions needs to be investigated further. This study considers the waveform competition between a multifunction radar and a transmit/receive time-sharing jammer. Two types of jamming schemes are used, suppression and deception, which markedly expand the current game space employed in radar reinforcement learning. A two-agent intelligent game framework is constructed to investigate control strategies for the training procedure under imperfect observation states. A heterogeneous reinforcement learning algorithm with an opponent sampling policy is proposed. A curiosity-driven exploration method is developed to increase the efficiency and stability of training. Simulation results showed that the proposed algorithm enables the radar agent to obtain higher rewards during the game. Furthermore, the opponent sampling policy prevented the radar from converging to local optimum solutions.